How a Poisson Model Predicts Football Scores

It is Saturday afternoon and two mid-table teams are about to kick off. A friend says confidently, "2-1 to the home side." Another says it will be a 0-0 bore. Who is right? Neither can know, but a simple mathematical tool can tell you how likely each scoreline really is. That tool is the Poisson model, and it sits behind a large share of football forecasting.

The basic idea: goals are rare events

Football is a low-scoring sport. Top leagues typically average somewhere around 2.5 to 3 goals per match, and most matches finish with only a handful of goals. Goals also arrive at fairly unpredictable moments, and one goal does not obviously cause the next.

Statisticians have a classic tool for counting rare, scattered events: the Poisson distribution. It was used for things like the number of calls arriving at a switchboard in an hour. Football goals share the same feel, which is why the model has been applied to the sport for decades.

What the Poisson distribution actually says

The distribution needs just one input: the average number of goals expected, usually written as lambda. Give it that number and it returns the probability of seeing exactly 0, 1, 2, 3 or more goals.

Some useful properties:

• If a team is expected to score about 1 goal, the most likely outcomes are 0 or 1 goal, with 2 or more far less common.
• If a team is expected to score about 2 goals, 1 and 2 goals become the most likely outcomes.
• The distribution is always skewed: there is a long tail of unlikely high scores, but zero is never impossible.

The model then treats the two teams as independent. It computes a distribution for the home team, another for the away team, and multiplies the two probabilities to get the chance of any exact scoreline.

Step 1: estimating expected goals for each team

Everything depends on the two lambdas. A common way to build them uses attack strength and defence strength:

• Attack strength: a team's goals scored per match divided by the league average.
• Defence strength: a team's goals conceded per match divided by the league average. A value below 1 means a better-than-average defence.
• Home advantage: home teams score more on average than away teams, so the model uses separate league averages for home and away goals.

The expected goals for the home team are then: league home average x home team attack x away team defence. The away team's figure uses the same logic in reverse. More advanced versions replace raw goals with expected goals (xG), which measures the quality of chances and is less noisy than actual results, and they give more weight to recent matches.

A worked example

The numbers below are an illustrative example, not real match data. Suppose the league averages 1.45 goals per match for home teams and 1.15 for away teams.

• Home team attack 1.15, away team defence 0.95: 1.45 x 1.15 x 0.95 is about 1.6 expected goals for the home side.
• Away team attack 0.95, home team defence 1.00: 1.15 x 0.95 x 1.00 is about 1.1 expected goals for the away side.

Feeding 1.6 and 1.1 into the Poisson formula gives these approximate probabilities for the number of goals each team scores:

To get a scoreline, multiply one cell from each column. For example, a 1-1 draw is 32.3% x 36.6%, which is about 11.8%. A 2-1 home win is 25.8% x 36.6%, about 9.5%. A 0-0 is 20.2% x 33.3%, about 6.7%.

Notice something important: even the single most likely scoreline here (1-1) only has a probability of around one in nine. No individual scoreline is ever very likely, which is why forecasting exact scores is so hard.

From scorelines to match outcomes and goal markets

The real power of the model appears when you add up many scorelines. Summing the right cells of the full grid gives:

• Home win: all scorelines where the home team scores more. In this example, about 49%.
• Draw: all scorelines with equal goals. About 25%.
• Away win: all scorelines where the away team scores more. About 26%.
• Over 2.5 goals: the total expected goals is 2.7, which gives roughly a 51% chance of three goals or more.
• Both teams to score: (1 - 20.2%) x (1 - 33.3%), about 53%.

So a single pair of expected-goal numbers produces a whole family of probabilities. This is how a model can say a team has a 49% chance of winning without claiming to know what will happen. AI Football Predictor publishes these kinds of numbers, including match probabilities and xG, for every match it covers.

Where the Poisson model falls short

The model is elegant, but it is a simplification. Be aware of its main weaknesses:

• Independence assumption: in reality, the two teams' scoring is not fully independent. A team that goes ahead may sit back, and a trailing team may push forward.
• Low scores: basic Poisson tends to misjudge the frequency of 0-0, 1-0, 0-1 and 1-1. Refinements such as the Dixon-Coles adjustment correct this.
• Variation in goals: real football often has slightly more variation than a pure Poisson model expects, with more very low and very high scoring games.
• Changing circumstances: injuries, red cards, rotation, fixture congestion and managerial changes are not captured by averages alone.
• Small samples: early in a season, attack and defence ratings are based on few matches and can be misleading.

Good models handle these issues by blending several seasons, weighting recent form, adjusting for opponent strength and using xG rather than only goals.

How to read the numbers sensibly

A probability is not a promise. A team with a 49% chance of winning will fail to win more often than not, and that is perfectly consistent with the model being correct. To judge a model fairly you need many matches: out of all the games it rated at 25% for a draw, roughly a quarter should end level.

Use the output to understand a match, not to predict it with certainty. Look at the gap between the two expected goals values, check whether the totals suggest an open or tight game, and treat close probabilities as genuinely open contests. If you are tempted to connect any statistic to gambling, remember that statistics are not guarantees, that gambling is for adults only and that it carries real financial and personal risks.

Key takeaways

• The Poisson model converts an expected number of goals into probabilities for 0, 1, 2 or more goals.
• Multiplying the two teams' distributions gives the probability of every exact scoreline.
• Adding scorelines together gives win, draw and loss probabilities and goal totals.
• The model is a useful, transparent baseline, but it has known limits and works best when fed quality inputs such as xG.